A forecast says a team has a 60% chance of winning. Does that mean it will win? Does it mean the model is 60% sure? Does it mean the other side has no chance? The answer to all three is no, and clearing up why is the point of this guide.
Football probability is not complicated, but it comes with a few ideas that are easy to misread. Here is how expected goals, win chances and model confidence fit together, with a worked example you can follow with a calculator.
Key Takeaways
- A probability describes how often something would happen over many similar matches, not what will happen in one.
- Win, draw and loss percentages add up to 100%, and the draw is a real outcome.
- Expected goals (xG) measure chance quality, and models can turn expected goals into scoreline probabilities.
- Confidence is not the same as probability: a forecast can be precise or uncertain, and well calibrated or not.
- A calibrated model’s 70% events should happen about 70% of the time over many matches.
Probability Basics: The 0 to 100 Scale
Probability runs from 0% (impossible) to 100% (certain). In football, almost everything sits in between. A 60% win chance means that if the same match were played many times under the same conditions, the team would win about 6 in 10.
That also means a 60% favourite loses or draws about 4 times in 10. This is why a single result rarely proves a forecast right or wrong.
Win, Draw and Loss Percentages
Most forecasts split a match into three outcomes: home win, draw and away win. The three always add up to 100%, so a card might read Home 49%, Draw 25% and Away 26%.
The draw is the outcome people forget. In top leagues roughly a quarter of matches finish level, so a 25% draw chance is not a quirk. It is simply what the sport looks like.
Expected Goals as an Input
Expected goals (xG) give each shot a value between 0 and 1 based on how likely that kind of chance is to become a goal. A penalty carries a high value and a long-range effort a low one. Add up the values over a match and you get a measure of the quality of chances created.
Over a season, xG often tells you more about a team’s real level than the goals scored alone, because goals include luck and finishing streaks. Read More: The xG explainer covers how expected goals are calculated and how to use them.
From Expected Goals to Scoreline Probabilities
Here is where it gets interesting. A model can take the expected goals for each side and turn them into the chance of every scoreline.
The classic approach uses a Poisson distribution, a standard way to describe how often a rare-ish event, like a goal, happens in a fixed period. The idea has been used for football since the early 1980s, and Dixon and Coles extended it in 1997 to reflect how teams’ form changes over time.
1. A Worked Example
Say the home side is expected to score 1.6 goals and the away side 1.1. Run those through the Poisson maths and you get the following scorelines, rounded to one decimal place.
| Scoreline | Probability |
|---|---|
| 1-1 | 11.8% |
| 1-0 | 10.8% |
| 2-1 | 9.5% |
| 2-0 | 8.6% |
| 0-1 | 7.4% |
| 0-0 | 6.7% |
2. Adding Up the Scorelines
Group the scorelines by result and the numbers become familiar. Adding every home-win scoreline gives roughly 49%, the draws about 25% and the away wins about 26%.
The same grid gives other markets. Both teams scoring comes out at about 53%, and three or more goals at about 51%.
3. A Useful Surprise
Look at the table again. The single most likely scoreline is 1-1 at 11.8%, yet the most likely result is a home win at about 49%. That is because the home win is spread across many scorelines, while the draw is concentrated in a few.
It is a good reminder that “most likely scoreline” and “most likely result” are different questions.
4. A Note on Simplicity
This example assumes the two teams score independently and that expected goals are known precisely. Real models adjust for both, but the logic is the same. Treat the example as a teaching tool, not a production forecast.
Home Advantage and Base Rates
Forecasts start from base rates. Home teams win more often than away teams in most leagues, and roughly a quarter of matches end level. A model then moves away from those starting points based on team strength, form and the other inputs it uses.
If a prediction card sits close to the base rates, the model sees little to separate the two sides. If it sits far from them, the model sees a clear difference in quality.
Why Single Matches Are Noisy
Football is a low-scoring sport, so one deflection or one refereeing call can flip a result. That randomness is why even a strong favourite does not win every time.
Take a team with a 70% chance in each of 10 separate matches. The chance it wins all 10 is 0.7 multiplied by itself 10 times, which is about 2.8%. So in more than 97% of such runs, at least one of the 10 will go the other way, and that is the forecast working as intended.
What Model Confidence Really Means
People say a model is “confident” when it gives a team 75% or more. But confidence in everyday speech mixes up three separate ideas.
1. Probability Versus Confidence
A probability is the model’s estimate of an outcome. Confidence is a statement about how much to trust that estimate. A 55% forecast based on strong data can deserve more trust than a 70% forecast based on very little.
2. Calibration
Calibration asks whether the percentages line up with reality. If a model says 70% for 100 different matches, about 70 of those events should happen. If only 55 do, the model is overconfident, and if 85 do, it is too cautious.
3. Sample Size and Uncertainty
A handful of matches cannot show calibration. A coin that lands heads 7 times in 10 flips is probably still a fair coin. The same logic applies to forecasts, which is why accuracy is judged over hundreds of matches rather than a single weekend.
How Forecasts Are Judged
Forecasters use scoring rules to measure how good a set of probabilities was. The Brier score averages the squared gap between the forecast probability and what actually happened, where 1 means it happened and 0 means it did not. Lower is better.
For football, where results are ordered (win, draw, loss), a rule called the ranked probability score was proposed by Constantinou and Fenton in 2012 as a better fit. The TuSportNews guide Are AI Football Predictions Accurate?, walks through these ideas in plain language.
Putting It All Together
Let’s join the pieces with the earlier example. A card shows Home 49%, Draw 25% and Away 26%, and the underlying expected goals are 1.6 for the home side and 1.1 for the visitors.
That tells you the home side is the more likely winner but far from a certainty, the draw is a live outcome and the most likely single scoreline is 1-1. It also tells you that if the home side loses, the model has not failed, because a 51% chance of not winning was always part of the picture.
Reading a Prediction Card
Putting it together, here is a simple routine for reading any card:
- Read all three percentages, not just the biggest one.
- Check the draw. It is rarely zero.
- Look at the gap between the top two outcomes, which tells you how open the match is.
- Remember the base rate. A 60% favourite fails 4 times in 10.
- Judge the model over time, not by one result.
The AI football predictions hub lists these cards for upcoming fixtures. For French football, the Ligue 1 predictions page and the Ligue 1 profile are a good place to practise, and the Champions League predictions page and Champions League profile cover Europe’s top club competition.
Read More: How to Use AI Football Predictions goes through a prediction card line by line.
Common Mistakes
These slips come up often:
- Reading 60% as “will win”. It means 6 in 10 over many matches.
- Ignoring the draw. It carries real probability in almost every match.
- Confusing confidence with probability. They answer different questions.
- Judging a model on one result. Calibration needs a large sample.
- Treating xG as a prediction. It measures chance quality, and models use it as an input.
Conclusion
Football probability comes down to a few ideas. Percentages describe long-run frequencies, the three result probabilities add up to 100%, and expected goals can be turned into scoreline probabilities with a bit of maths. Confidence is about trust in the estimate, which is best judged by calibration over many matches.
With those basics in place, any forecast becomes easier to read and harder to over-trust. Look at all three numbers, remember the draw, and judge a model by its record, not by a single Saturday. To compare a model with the market, learn to convert odds to implied probability and read odds comparison versus AI predictions.
This article is for educational purposes only and does not encourage gambling. Readers must be 18 or over, and anyone who feels gambling is affecting their life should contact a local support service.